| Rainfall forecast is closely related to people’s production and life,especially the short-term approaching rainfall forecast.The short-term approaching rainfall forecast mainly predicts the rainfall in the next few hours through radar echoes.Accurate prediction of short-term approach can effectively deal with major rainfall disasters,effectively protect people’s life safety,and avoid major economic losses.Therefore,accurate prediction of short-term impending rainfall has always been one of the important research directions in the field of meteorology.In recent years,with the continuous development of machine learning,especially deep learning,artificial intelligence methods have been widely used in various fields.The same meteorological field has been combined with artificial intelligence methods to achieve more accurate short-term imminent rainfall forecasts.The main work of this paper is to study the deep learning method based on ConvLSTM to achieve more accurate prediction of short-term approaching rainfall.This paper mainly does the following aspects.A new radar echo sequence prediction model PatchConvLSTM is proposed,Mainly by improving the original ConvLSTM and inputting a large-scale radar echo map,the model can learn relevant information outside the prediction area,and process it by dividing into Patchblocks.That is,the important information of the original input is retained and the computing resources are saved,so as to realize the accurate prediction of the radar echo in the central area.After that,the attention mechanism is further introduced to improve the model’s ability to extract spatial features and further improve the model prediction accuracy.A neural network method is proposed to realize the quantitative estimation of radar reflectivity for rainfall.The reflectivity factor in the radar echo image and the rainfallrelated parameters detected by the ground rain gauge station are mainly extracted through the time matching algorithm to form a usable data set.Design the neural network model to make its estimation accuracy better than the traditional Z-R relationship.Optimizing the backpropagation algorithm by data augmentation and using a weighted loss function.Improved neural network to increase its accuracy in estimating larger rainfall categories.Through ablation experiments,the reflectivity is analyzed to estimate the accuracy of rainfall.Designing a short-term near rainfall prediction system and applying it to the dataset of Heilong Province.First,cleaning and processing of various radar sites in Heilongjiang Province.Extract raw radar echo data using projection mapping algorithm,using KNearest Neighbors Algorithm to Repair Data.The data sets of each radar site are stitched together by a spatio-temporal matching algorithm.resulting in a usable large-scale radar echo dataset.Fine-tune the optimal radar echo prediction model on this dataset to make the model suitable for Heilongjiang Province data.A set of short-term near rainfall prediction system is designed by integrating the method of this paper,and the QT interface is designed to realize its engineering application in some areas of Heilongjiang Province.The above is the main content of this paper,and by improving the existing model to achieve more accurate prediction of radar echo,using the neural network method to estimate rainfall through the predicted radar echo,so as to achieve accurate prediction of short-term imminent rainfall,And apply it to some areas of Heilong Province. |
